Open Source High Content Analysis Utilizing Automated Fluorescence Lifetime Imaging Microscopy

The increasing use of automated high content analysis (HCA) in drug discovery¹ to assay compound libraries is complemented by the trend in life science basic research towards automated imaging experiments for systematic studies, e.g., for phenotypic screens of siRNA libraries. Automated HCA is also becoming increasingly important for smaller scale biological studies that have typically been implemented with manual experiments with tens of dishes of cells imaged with conventional microscopes. The statistical power conferred by the ability to assay and analyze hundreds to thousands of fields of view enables averaging of experimental noise (including "biological noise") and the automation of image data acquisition and analysis of large sample arrays can help eliminate operator bias and often may be utilized to detect the occurrence of systematic errors, such as a change in the IRF profile during an acquisition. Fluorescence based assays are widely implemented in HCA, with most commercially available HCA instrumentation featuring fluorescence intensity imaging in one or more spectral channels to map the relative distribution and co-localization of labeled proteins. More sophisticated fluorescence imaging modalities, such as fluorescence lifetime imaging (FLIM), polarization anisotropy imaging and time-resolved fluorescence anisotropy imaging, are not widely exploited in commercial HCA instruments, although they offer powerful quantitative readouts, e.g., of protein interactions. Here we present an open source approach to implementing FLIM in an automated microscope for HCA.

Fluorescence lifetime provides an inherently ratiometric spectroscopic readout that can be used to distinguish different molecular species or to probe the local molecular environment of a fluorophore and is insensitive to fluorophore concentration, to excitation/detection efficiencies, and to the impact of scattering and sample absorption². Furthermore, since fluorescence lifetime measurements can be made in a single spectral channel, they are directly comparable between different instruments and samples without calibration. They can be applied to endogenous fluorophores, including some cellular metabolites, lipids and extracellular matrix components, to provide intrinsic readouts of biological processes and pathologies. Thus label-free FLIM can map metabolic changes in cells^3,4 and has been applied to monitor stem cell differentiation^5,6 and the growth of collagen scaffolds⁷. We recently demonstrated that FLIM HCA can be used for automated label-free assays of cellular metabolism utilizing autofluorescence⁸. More commonly, however, fluorescence lifetime measurements and FLIM are applied to exogenous fluorophores that label specific biomolecules, often implemented using dyes that stain cellular compartments, using dyes coupled to antibodies that target specific proteins in fixed cells, or using genetically expressed fluorophores coupled to specific proteins. Fluorescence lifetime can be used to provide robust quantitative readouts of fluorescent probes sensing their physical or chemical environment. For examples, dyes have been deployed to sense the local lipid phase of cell membranes⁹ or cell viscosity¹⁰ and genetically expressed fluorescent protein-based biosensors have been developed to report concentrations of cell signaling molecules including as IP3¹¹, PIP2¹² and calpain¹³ or ions such as calcium^14,15, potassium¹⁶ and chloride¹⁷.

Many such biosensors utilize Förster Resonance Energy Transfer (FRET)^18,19 to provide readouts of changes in the biosensor conformation. FRET between "donor" and "acceptor" fluorophores occurs via a short range direct dipole-dipole interaction for which fluorophores are typically separated by less than ~10 nm. Thus the detection of FRET indicates the close proximity of appropriately labeled proteins and hence provides a means to read out protein-protein interactions²⁰, such as binding or oligomerization — either as end points in fixed cells or as dynamic events in time course measurements of live cells. By labeling an appropriate molecular construct with both donor and acceptor fluorophores, it is also possible to read out conformational changes — or cleavage — of the construct via the change in FRET and this is the basis of many biosensors²¹. Measurements of the FRET efficiency can provide information about the donor-acceptor distance and the population fraction of donor fluorophores undergoing FRET. Thus, FRET-based imaging techniques can be used to map and quantitate molecular interactions in space and time, e.g., to elucidate cell signaling processes²². Automating such imaging techniques could provide high throughput assays based on readouts of signaling pathways or networks.

In HCA, the FRET readout most commonly implemented is spectral ratiometric imaging, where changes in the ratio of acceptor to donor intensity are mapped. While this approach is readily implemented — and is available in many commercially available multiwell plate readers — quantitative spectrally resolved FRET measurements require calibration of the spectral response of the system²³. This includes spectral cross-talk arising from spectral bleed-through and direct excitation of the acceptor as well as the transmission characteristics of the instrument and the sample (inner filter effect). In principle, this entails measurement of additional "control" samples that are labeled with either donor-only or acceptor-only.

From such spectral ratiometric FRET measurements, an effective FRET efficiency (the product of the actual FRET efficiency and fraction of FRETing donor/acceptor molecules) can be obtained along with the relative proportion of donor and acceptor molecules²⁴. Spectral ratiometric FRET is often used with unimolecular FRET biosensors, where the donor/acceptor stoichiometry can be assumed to be known. To determine the population fractions of the FRETing donor and acceptor, e.g., to obtain a dose response curve, knowledge of the actual FRET efficiency is required. This can be obtained by measuring a positive FRET control sample with known properties or by using a different method to measure the FRET efficiency, such as acceptor photobleaching or FLIM.

Using fluorescence lifetime readouts, FRET can be detected and quantified by measurements of the donor emission alone — removing the need for spectral calibration and further control samples. Thus, FLIM FRET readouts can be compared between instruments and between different samples — allowing data from endoscopy or tomography of live disease models²⁵ to be benchmarked against measurements of cell-based assays. By fitting donor fluorescence decay profiles to an appropriate multi-exponential decay model corresponding to FRETing and non-FRETing donor populations, FRET efficiencies and population fractions can, in principle, be obtained directly without further measurements. These significant advantages, however, come at the cost of FLIM requiring more detected photons per pixel than spectrally resolved intensity imaging — typically hundreds of photons are required to achieve an accuracy in lifetime determination of 10% when fitting to a single exponential decay model and when fitting fluorescence decay profiles to complex (multiexponential decay) models, the number of detected photons required increases to thousands. This has conventionally led to long acquisition times (tens to hundreds of seconds per field of view) for FLIM, particularly when using time correlated single photon counting (TCSPC) implemented with sequential pixel acquisition in laser scanning microscopes. Attempts to increase the imaging speed by using higher excitation intensities can lead to significant photobleaching and/or phototoxicity. Together with a lack of appropriate commercially available instrumentation and software tools, this has hindered FLIM from being utilized in HCA for drug discovery or systems biology, where hundreds to thousands of fields of view are to be imaged. Laser scanning TCSPC FLIM has been implemented in an automated multiwell plate microscope²⁶ for secondary measurements following identification of "hits" by steady-state polarization-resolved fluorescence anisotropy imaging²⁷, which can reach similar imaging speeds to spectral ratiometric imaging²⁸ with which it shares many advantages and disadvantages. Fluorescence anisotropy imaging exploits the decrease in fluorescence anisotropy of the acceptor in the presence of FRET and is also sensitive to spectral cross-talk (e.g., direct excitation of the acceptor). It requires calibration to account for polarization cross talk and does not provide a direct measurement of FRET efficiency.

To realize the advantages of FLIM for HCA, we have worked to address the issues of speed of image acquisition and wider access to the technology. Here, we provide a complete list of open source software and hardware components that can be utilized to implement the automated rapid FLIM multiwell plate reader that is described below, together with exemplar FRET based-assays. In our approach²⁹, FLIM is realized using wide-field time-gated imaging using a gated optical image intensifier (GOI), which is illustrated in Figure 1 that can provide faster imaging rates and lower photobleaching than single-beam scanning TCSPC due to the parallel pixel interrogation. Our openFLIM-HCA instrument can provide unsupervised FLIM, requiring only a few seconds to acquire FLIM data detecting a few 100 photons per pixel (depending on the brightness of the sample). Combined with the time required to move the sample from the previous field of view, to adjust the acquisition settings and to maintain the sample in focus, this typically results in multiwell plate acquisition times of ~10 s per field of view^25,28. Optical sectioning can be implemented using a quasi-wide-field Nipkow ("spinning disc") scanner³⁰.

For FLIM data analysis, we have developed an open source software tool called FLIMfit³¹ that is able to rapidly analyze multiwell plate FLIM datasets on conventional PC workstations and is a plug-in for OMERO³². FLIMfit provides global analysis capabilities (discussed in Section 1.2) that enable FLIM data to be fitted to complex models with only a few 100 photons per pixel under the assumption that the fluorescence lifetime components are invariant across an image or region of interest (ROI), therefore reducing the number of detected photons required, the light exposure to the sample and image acquisition time. Running FLIMfit on a standard desktop PC, requires only 10s of seconds of computational time to analyze data sets comprising hundreds of FOVs. Thus both FLIM data acquisition and analysis can be undertaken on timescales that are practical for HCA.

openFLIM-HCA hardware

Our implementation of FLIM HCA comprises six key components: (i) a fluorescence microscope frame with optical autofocus; (ii) a motorized (x-y-z) microscope stage; (iii) an excitation laser source; (iv) a wide-field time-gated FLIM system with gated optical intensifier (GOI), delay generator and camera; (v) a computer for data acquisition and analysis and (vi) a Nipkow disc scanner unit for optically sectioned imaging to suppress out of focus light. This can be important when imaging weak signals, e.g., from FRET biosensors at the cell plasma membrane, that could be swamped by contributions from cytoplasmic fluorescence or background from coverslips or culture medium. Time-gated FLIM data is acquired by illuminating the sample with the ultrashort pulsed excitation radiation, imaging the fluorescence emission to the photocathode of the GOI and acquiring a series of CCD images of the phosphor of the GOI for a range of settings of time delay of the optical intensifier gate after the arrival of the excitation pulses. As the time gate delay following excitation is increased, the fluorescence signal is seen to decrease and the series of time-gated fluorescence intensity images acquired on the CCD form the data set required to analyze the decay profiles of all the fluorophores in the field of view. The gating of the GOI is synchronized to the excitation pulses and, for the series of CCD acquisitions, the delay of the time gate relative to the excitation pulses is set to a series of increasing values over the desired range. This synchronization is achieved using the delay generator that comprises a "fast delay box" (providing delays up to 20 ns in 1 ps steps) and a "coarse delay box" that provides delays with a minimum step of 25 ps.

For FLIM, excitation can be provided by any ultrafast laser. For convenience, we typically employ a fiber-laser-pumped supercontinuum source that provides picosecond pulses tunable across the visible spectrum from which the appropriate excitation radiation can be selected for any fluorophore using a motorized filter wheel with different band pass filters. Typically, we use an average power of ~100-200 µW at the sample with pulses at 40 or 60 MHz repetition rate. Such excitation powers could also be provided by ultrafast laser sources based on frequency doubling of mode-locked titanium doped sapphire lasers. Note that an excitation pulse duration below ~100 ps is sufficient for most experiments. A second motorized filter wheel equipped with neutral density filters is used to regulate the excitation intensity. For safety and convenience, the excitation radiation may be delivered via a single mode optical fiber. The configurations for wide-field FLIM and optically sectioned FLIM are presented in Figures 2 and 3 respectively. For more details of the precise components used, please visit the openFLIM-HCA wiki³³. A copy of the current parts list is given in the Supplementary Material.

For wide-field FLIM, the excitation radiation is required to illuminate the whole field of view as uniformly as possible. For this we direct the excitation laser beam to a holographic "top-hat" diffuser that is rotated at 10-50 Hz to time-average the laser speckle, with the plane of the diffuser being imaged to the back focal plane of the objective microscope lens. The fluorescence image from the output port of the microscope is relayed onto the photocathode of the GOI and the phosphor of the GOI (active area diagonal 18 mm) is imaged onto the sensor of a cooled scientific CCD (chip size 10.2 mm x 8.3 mm) camera using a pair of camera lenses providing a magnification of 0.7. The system is controlled by a standard PC that is interfaced to the instrumentation using RS232 protocols. In addition, an Arduino microprocessor is used to control a shutter in the excitation light path that is closed except when the CCD camera is acquiring FLIM data and to record the signal from a photodiode that is used to measure the average power from the spectrally filtered supercontinuum excitation source. For optically sectioned FLIM, the excitation laser radiation is coupled into a single-mode polarization-maintaining optical fiber that connects directly to the Nipkow disc scanner unit, as shown in Figure 3. The output fluorescence image from the Nipkow scanner unit is relayed onto the photocathode of the GOI and the rest of the system is the same as for wide-field FLIM.

openFLIM-HCA software

The data acquisition software is written in µManager³⁴. It provides full HCA data acquisition functionality including options to change the sequence in which the wells of the plate are imaged, to acquire time courses for any FOV, to automatically maintain focus during measurements and to control the excitation and emission filter wheels and the dichroic changer for automated multicolor imaging. It is also possible to run a "prescan" acquisition of fluorescence intensity in order to automatically identify and record the positions of suitable FOVs for a subsequent FLIM acquisition according to criteria, e.g., based on intensity. FLIM HCA data can be saved as a series of TIFF files, as a series of OME-TIFF files (i.e., one per FOV) or as a single OME-TIFF file incorporating all the data from a multiwell plate. More information and a detailed description of the µManager software can be found on the openFLIM-HCA wiki³³.

The data analysis software, FLIMfit³¹, an open source MATLAB-based client for the OMERO image data management platform³², is able to read FLIM data from an OMERO server or from computer disk, as indicated in Figure 4. It has been designed to analyze data from TCSPC or wide-field time-gated FLIM instruments and provides many capabilities to fit data from openFLIM-HCA including fitting to monoexponential decay profiles, and to double exponential and higher complexity decay profiles, including models such as polarization-resolved fluorescence decay profiles. It can also take account of fixed and time-varying background signals and can utilize a measured spatio-temporal instrument response function (IRF) or can fit using an idealized IRF that may be time-shifted on a pixel-wise basis by an amount determined from the full experimental IRF measurement. A global time difference (t₀) between the IRF and a fluorescent sample can be determined from a measurement of a fluorescence standard. Spatial variations in background signals and IRF can also be taken into account. FLIMfit is able to fit FLIM data on a pixel-wise basis or using "global binning" or "global fitting" to facilitate fitting of FLIM data with modest (hundreds) photon numbers to complex decay models.

Global binning entails aggregating all the detected photons from the pixels within a user-defined ROI at each time point to provide sufficient photon numbers to enable fitting to complex decay models. The ROI can be the entire field of view (FOV), it can be manually defined by the user, or it can be determined using image segmentation tools. The ROI can also be defined using masks imported into FLIMfit from other image segmentation software. For FRET applications, it is common to use global binning to fit to a double exponential decay model to determine the fluorescence lifetime components corresponding to FRETing and non-FRETing donor fluorophores that are assumed to be spatially invariant across the ROI and then to refit on a pixel-wise basis with these lifetime component vales fixed in order to obtain the FRETing donor population fraction in each pixel.

Global fitting entails simultaneously fitting the lifetime components and pixel-wise FRETing donor population fractions across a whole FOV- or across a FLIM data set that could comprise multiple FOVs, e.g., from a multiwell plate experiment or a time series of fluorescence lifetime images. Global fitting is able to exploit the spatial variation of population fractions across the image that is lost in the global binning approach to provide stronger constraints on the component lifetime estimation — and therefore more reliable results — albeit at the cost of significantly more computation³⁵. FLIMfit can rapidly analyze multiwell plate FLIM datasets with global fitting on conventional PC workstations with, for example, a multiwell time-gated FLIM data set, acquired with five time gates for each of 394 FOV each comprising 672 x 512 pixels, requiring only 32 seconds and 2 GB of memory to fit to a double exponential decay model³¹. This has been achieved by utilizing a separable nonlinear least square fitting algorithm implemented using multithreaded parallel algorithms to enable effective scaling on multicore processors and the FLIMfit code has been optimized to minimize memory usage. Figure 5 shows a snapshot of the graphical user interface of FLIMfit and more details can be found at the web page given in reference 36. Further information on global fitting and global binning can be found in reference 31.

The following protocol for FLIM HCA using our openFLIM-HCA instrumentation and software can be adapted for a wide range of cell imaging experiments with FLIM readouts. In general, multiwell plate FRET experiments should be designed to include replicate wells of positive and negative biological controls in addition to the conditions being studied. Here, we describe the specific implementation to perform optically sectioned FLIM (see Figure 3) of Cos cells expressing FRET constructs consisting of the mTurquoise fluorescent protein and Venus fluorescent protein joined by a short peptide linker. Here the mTq32V, mTq17V and mTq5V FRET constructs (discussed in Representative Results section) provide low, medium and high FRET efficiencies together with the non-FRETing mTq5A construct. These constructs and the other materials required to follow this protocol are listed in the Materials List.